| 1 | //===- AArch64RegisterInfo.cpp - AArch64 Register Information -------------===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This file contains the AArch64 implementation of the TargetRegisterInfo |
| 10 | // class. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #include "AArch64RegisterInfo.h" |
| 15 | #include "AArch64FrameLowering.h" |
| 16 | #include "AArch64InstrInfo.h" |
| 17 | #include "AArch64MachineFunctionInfo.h" |
| 18 | #include "AArch64SMEAttributes.h" |
| 19 | #include "AArch64Subtarget.h" |
| 20 | #include "MCTargetDesc/AArch64AddressingModes.h" |
| 21 | #include "MCTargetDesc/AArch64InstPrinter.h" |
| 22 | #include "llvm/ADT/BitVector.h" |
| 23 | #include "llvm/BinaryFormat/Dwarf.h" |
| 24 | #include "llvm/CodeGen/LiveRegMatrix.h" |
| 25 | #include "llvm/CodeGen/MachineFrameInfo.h" |
| 26 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
| 27 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
| 28 | #include "llvm/CodeGen/RegisterScavenging.h" |
| 29 | #include "llvm/CodeGen/TargetFrameLowering.h" |
| 30 | #include "llvm/IR/DebugInfoMetadata.h" |
| 31 | #include "llvm/IR/DiagnosticInfo.h" |
| 32 | #include "llvm/IR/Function.h" |
| 33 | #include "llvm/Target/TargetOptions.h" |
| 34 | #include "llvm/TargetParser/Triple.h" |
| 35 | |
| 36 | using namespace llvm; |
| 37 | |
| 38 | #define GET_CC_REGISTER_LISTS |
| 39 | #include "AArch64GenCallingConv.inc" |
| 40 | #define GET_REGINFO_TARGET_DESC |
| 41 | #include "AArch64GenRegisterInfo.inc" |
| 42 | |
| 43 | AArch64RegisterInfo::AArch64RegisterInfo(const Triple &TT, unsigned HwMode) |
| 44 | : AArch64GenRegisterInfo(AArch64::LR, 0, 0, 0, HwMode), TT(TT) { |
| 45 | AArch64_MC::initLLVMToCVRegMapping(MRI: this); |
| 46 | } |
| 47 | |
| 48 | /// Return whether the register needs a CFI entry. Not all unwinders may know |
| 49 | /// about SVE registers, so we assume the lowest common denominator, i.e. the |
| 50 | /// callee-saves required by the base ABI. For the SVE registers z8-z15 only the |
| 51 | /// lower 64-bits (d8-d15) need to be saved. The lower 64-bits subreg is |
| 52 | /// returned in \p RegToUseForCFI. |
| 53 | bool AArch64RegisterInfo::regNeedsCFI(MCRegister Reg, |
| 54 | MCRegister &RegToUseForCFI) const { |
| 55 | if (AArch64::PPRRegClass.contains(Reg)) |
| 56 | return false; |
| 57 | |
| 58 | if (AArch64::ZPRRegClass.contains(Reg)) { |
| 59 | RegToUseForCFI = getSubReg(Reg, Idx: AArch64::dsub); |
| 60 | for (int I = 0; CSR_AArch64_AAPCS_SaveList[I]; ++I) { |
| 61 | if (CSR_AArch64_AAPCS_SaveList[I] == RegToUseForCFI) |
| 62 | return true; |
| 63 | } |
| 64 | return false; |
| 65 | } |
| 66 | |
| 67 | RegToUseForCFI = Reg; |
| 68 | return true; |
| 69 | } |
| 70 | |
| 71 | const MCPhysReg * |
| 72 | AArch64RegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const { |
| 73 | assert(MF && "Invalid MachineFunction pointer." ); |
| 74 | |
| 75 | auto &AFI = *MF->getInfo<AArch64FunctionInfo>(); |
| 76 | const auto &F = MF->getFunction(); |
| 77 | const auto *TLI = MF->getSubtarget<AArch64Subtarget>().getTargetLowering(); |
| 78 | const bool Darwin = MF->getSubtarget<AArch64Subtarget>().isTargetDarwin(); |
| 79 | const bool Windows = MF->getSubtarget<AArch64Subtarget>().isTargetWindows(); |
| 80 | |
| 81 | if (TLI->supportSwiftError() && |
| 82 | F.getAttributes().hasAttrSomewhere(Kind: Attribute::SwiftError)) { |
| 83 | if (Darwin) |
| 84 | return CSR_Darwin_AArch64_AAPCS_SwiftError_SaveList; |
| 85 | if (Windows) |
| 86 | return CSR_Win_AArch64_AAPCS_SwiftError_SaveList; |
| 87 | return CSR_AArch64_AAPCS_SwiftError_SaveList; |
| 88 | } |
| 89 | |
| 90 | switch (F.getCallingConv()) { |
| 91 | case CallingConv::GHC: |
| 92 | // GHC set of callee saved regs is empty as all those regs are |
| 93 | // used for passing STG regs around |
| 94 | return CSR_AArch64_NoRegs_SaveList; |
| 95 | |
| 96 | case CallingConv::PreserveNone: |
| 97 | // FIXME: Windows likely need this to be altered for properly unwinding. |
| 98 | return CSR_AArch64_NoneRegs_SaveList; |
| 99 | |
| 100 | case CallingConv::AnyReg: |
| 101 | return CSR_AArch64_AllRegs_SaveList; |
| 102 | |
| 103 | case CallingConv::ARM64EC_Thunk_X64: |
| 104 | return CSR_Win_AArch64_Arm64EC_Thunk_SaveList; |
| 105 | |
| 106 | case CallingConv::PreserveMost: |
| 107 | if (Darwin) |
| 108 | return CSR_Darwin_AArch64_RT_MostRegs_SaveList; |
| 109 | if (Windows) |
| 110 | return CSR_Win_AArch64_RT_MostRegs_SaveList; |
| 111 | return CSR_AArch64_RT_MostRegs_SaveList; |
| 112 | |
| 113 | case CallingConv::PreserveAll: |
| 114 | if (Darwin) |
| 115 | return CSR_Darwin_AArch64_RT_AllRegs_SaveList; |
| 116 | if (Windows) |
| 117 | return CSR_Win_AArch64_RT_AllRegs_SaveList; |
| 118 | return CSR_AArch64_RT_AllRegs_SaveList; |
| 119 | |
| 120 | case CallingConv::CFGuard_Check: |
| 121 | if (Darwin) |
| 122 | report_fatal_error( |
| 123 | reason: "Calling convention CFGuard_Check is unsupported on Darwin." ); |
| 124 | return CSR_Win_AArch64_CFGuard_Check_SaveList; |
| 125 | |
| 126 | case CallingConv::SwiftTail: |
| 127 | if (Darwin) |
| 128 | return CSR_Darwin_AArch64_AAPCS_SwiftTail_SaveList; |
| 129 | if (Windows) |
| 130 | return CSR_Win_AArch64_AAPCS_SwiftTail_SaveList; |
| 131 | return CSR_AArch64_AAPCS_SwiftTail_SaveList; |
| 132 | |
| 133 | case CallingConv::AArch64_VectorCall: |
| 134 | if (Darwin) |
| 135 | return CSR_Darwin_AArch64_AAVPCS_SaveList; |
| 136 | if (Windows) |
| 137 | return CSR_Win_AArch64_AAVPCS_SaveList; |
| 138 | return CSR_AArch64_AAVPCS_SaveList; |
| 139 | |
| 140 | case CallingConv::AArch64_SVE_VectorCall: |
| 141 | if (Darwin) |
| 142 | report_fatal_error( |
| 143 | reason: "Calling convention SVE_VectorCall is unsupported on Darwin." ); |
| 144 | if (Windows) |
| 145 | return CSR_Win_AArch64_SVE_AAPCS_SaveList; |
| 146 | return CSR_AArch64_SVE_AAPCS_SaveList; |
| 147 | |
| 148 | case CallingConv::AArch64_SME_ABI_Support_Routines_PreserveMost_From_X0: |
| 149 | report_fatal_error( |
| 150 | reason: "Calling convention " |
| 151 | "AArch64_SME_ABI_Support_Routines_PreserveMost_From_X0 is only " |
| 152 | "supported to improve calls to SME ACLE save/restore/disable-za " |
| 153 | "functions, and is not intended to be used beyond that scope." ); |
| 154 | |
| 155 | case CallingConv::AArch64_SME_ABI_Support_Routines_PreserveMost_From_X1: |
| 156 | report_fatal_error( |
| 157 | reason: "Calling convention " |
| 158 | "AArch64_SME_ABI_Support_Routines_PreserveMost_From_X1 is " |
| 159 | "only supported to improve calls to SME ACLE __arm_get_current_vg " |
| 160 | "function, and is not intended to be used beyond that scope." ); |
| 161 | |
| 162 | case CallingConv::AArch64_SME_ABI_Support_Routines_PreserveMost_From_X2: |
| 163 | report_fatal_error( |
| 164 | reason: "Calling convention " |
| 165 | "AArch64_SME_ABI_Support_Routines_PreserveMost_From_X2 is " |
| 166 | "only supported to improve calls to SME ACLE __arm_sme_state " |
| 167 | "and is not intended to be used beyond that scope." ); |
| 168 | |
| 169 | case CallingConv::Win64: |
| 170 | if (Darwin) |
| 171 | return CSR_Darwin_AArch64_AAPCS_Win64_SaveList; |
| 172 | if (Windows) |
| 173 | return CSR_Win_AArch64_AAPCS_SaveList; |
| 174 | return CSR_AArch64_AAPCS_X18_SaveList; |
| 175 | |
| 176 | case CallingConv::CXX_FAST_TLS: |
| 177 | if (Darwin) |
| 178 | return AFI.isSplitCSR() ? CSR_Darwin_AArch64_CXX_TLS_PE_SaveList |
| 179 | : CSR_Darwin_AArch64_CXX_TLS_SaveList; |
| 180 | // FIXME: this likely should be a `report_fatal_error` condition, however, |
| 181 | // that would be a departure from the previously implemented behaviour. |
| 182 | LLVM_FALLTHROUGH; |
| 183 | |
| 184 | default: |
| 185 | if (Darwin) |
| 186 | return AFI.hasSVE_AAPCS(MF: *MF) ? CSR_Darwin_AArch64_SVE_AAPCS_SaveList |
| 187 | : CSR_Darwin_AArch64_AAPCS_SaveList; |
| 188 | if (Windows) |
| 189 | return AFI.hasSVE_AAPCS(MF: *MF) ? CSR_Win_AArch64_SVE_AAPCS_SaveList |
| 190 | : CSR_Win_AArch64_AAPCS_SaveList; |
| 191 | return AFI.hasSVE_AAPCS(MF: *MF) ? CSR_AArch64_SVE_AAPCS_SaveList |
| 192 | : CSR_AArch64_AAPCS_SaveList; |
| 193 | } |
| 194 | } |
| 195 | |
| 196 | const MCPhysReg *AArch64RegisterInfo::getCalleeSavedRegsViaCopy( |
| 197 | const MachineFunction *MF) const { |
| 198 | assert(MF && "Invalid MachineFunction pointer." ); |
| 199 | if (MF->getFunction().getCallingConv() == CallingConv::CXX_FAST_TLS && |
| 200 | MF->getInfo<AArch64FunctionInfo>()->isSplitCSR()) |
| 201 | return CSR_Darwin_AArch64_CXX_TLS_ViaCopy_SaveList; |
| 202 | return nullptr; |
| 203 | } |
| 204 | |
| 205 | void AArch64RegisterInfo::UpdateCustomCalleeSavedRegs( |
| 206 | MachineFunction &MF) const { |
| 207 | const MCPhysReg *CSRs = getCalleeSavedRegs(MF: &MF); |
| 208 | SmallVector<MCPhysReg, 32> UpdatedCSRs; |
| 209 | for (const MCPhysReg *I = CSRs; *I; ++I) |
| 210 | UpdatedCSRs.push_back(Elt: *I); |
| 211 | |
| 212 | for (size_t i = 0; i < AArch64::GPR64commonRegClass.getNumRegs(); ++i) { |
| 213 | if (MF.getSubtarget<AArch64Subtarget>().isXRegCustomCalleeSaved(i)) { |
| 214 | UpdatedCSRs.push_back(Elt: AArch64::GPR64commonRegClass.getRegister(i)); |
| 215 | } |
| 216 | } |
| 217 | // Register lists are zero-terminated. |
| 218 | UpdatedCSRs.push_back(Elt: 0); |
| 219 | MF.getRegInfo().setCalleeSavedRegs(UpdatedCSRs); |
| 220 | } |
| 221 | |
| 222 | const TargetRegisterClass * |
| 223 | AArch64RegisterInfo::getSubClassWithSubReg(const TargetRegisterClass *RC, |
| 224 | unsigned Idx) const { |
| 225 | // edge case for GPR/FPR register classes |
| 226 | if (RC == &AArch64::GPR32allRegClass && Idx == AArch64::hsub) |
| 227 | return &AArch64::FPR32RegClass; |
| 228 | else if (RC == &AArch64::GPR64allRegClass && Idx == AArch64::hsub) |
| 229 | return &AArch64::FPR64RegClass; |
| 230 | |
| 231 | // Forward to TableGen's default version. |
| 232 | return AArch64GenRegisterInfo::getSubClassWithSubReg(RC, Idx); |
| 233 | } |
| 234 | |
| 235 | const uint32_t * |
| 236 | AArch64RegisterInfo::getDarwinCallPreservedMask(const MachineFunction &MF, |
| 237 | CallingConv::ID CC) const { |
| 238 | assert(MF.getSubtarget<AArch64Subtarget>().isTargetDarwin() && |
| 239 | "Invalid subtarget for getDarwinCallPreservedMask" ); |
| 240 | |
| 241 | if (CC == CallingConv::CXX_FAST_TLS) |
| 242 | return CSR_Darwin_AArch64_CXX_TLS_RegMask; |
| 243 | if (CC == CallingConv::AArch64_VectorCall) |
| 244 | return CSR_Darwin_AArch64_AAVPCS_RegMask; |
| 245 | if (CC == CallingConv::AArch64_SVE_VectorCall) |
| 246 | return CSR_Darwin_AArch64_SVE_AAPCS_RegMask; |
| 247 | if (CC == CallingConv::AArch64_SME_ABI_Support_Routines_PreserveMost_From_X0) |
| 248 | return CSR_AArch64_SME_ABI_Support_Routines_PreserveMost_From_X0_RegMask; |
| 249 | if (CC == CallingConv::AArch64_SME_ABI_Support_Routines_PreserveMost_From_X1) |
| 250 | return CSR_AArch64_SME_ABI_Support_Routines_PreserveMost_From_X1_RegMask; |
| 251 | if (CC == CallingConv::AArch64_SME_ABI_Support_Routines_PreserveMost_From_X2) |
| 252 | return CSR_AArch64_SME_ABI_Support_Routines_PreserveMost_From_X2_RegMask; |
| 253 | if (CC == CallingConv::CFGuard_Check) |
| 254 | report_fatal_error( |
| 255 | reason: "Calling convention CFGuard_Check is unsupported on Darwin." ); |
| 256 | if (MF.getSubtarget<AArch64Subtarget>() |
| 257 | .getTargetLowering() |
| 258 | ->supportSwiftError() && |
| 259 | MF.getFunction().getAttributes().hasAttrSomewhere(Kind: Attribute::SwiftError)) |
| 260 | return CSR_Darwin_AArch64_AAPCS_SwiftError_RegMask; |
| 261 | if (CC == CallingConv::SwiftTail) |
| 262 | return CSR_Darwin_AArch64_AAPCS_SwiftTail_RegMask; |
| 263 | if (CC == CallingConv::PreserveMost) |
| 264 | return CSR_Darwin_AArch64_RT_MostRegs_RegMask; |
| 265 | if (CC == CallingConv::PreserveAll) |
| 266 | return CSR_Darwin_AArch64_RT_AllRegs_RegMask; |
| 267 | return CSR_Darwin_AArch64_AAPCS_RegMask; |
| 268 | } |
| 269 | |
| 270 | const uint32_t * |
| 271 | AArch64RegisterInfo::getCallPreservedMask(const MachineFunction &MF, |
| 272 | CallingConv::ID CC) const { |
| 273 | bool SCS = MF.getFunction().hasFnAttribute(Kind: Attribute::ShadowCallStack); |
| 274 | if (CC == CallingConv::GHC) |
| 275 | // This is academic because all GHC calls are (supposed to be) tail calls |
| 276 | return SCS ? CSR_AArch64_NoRegs_SCS_RegMask : CSR_AArch64_NoRegs_RegMask; |
| 277 | if (CC == CallingConv::PreserveNone) |
| 278 | return SCS ? CSR_AArch64_NoneRegs_SCS_RegMask |
| 279 | : CSR_AArch64_NoneRegs_RegMask; |
| 280 | if (CC == CallingConv::AnyReg) |
| 281 | return SCS ? CSR_AArch64_AllRegs_SCS_RegMask : CSR_AArch64_AllRegs_RegMask; |
| 282 | |
| 283 | // All the following calling conventions are handled differently on Darwin. |
| 284 | if (MF.getSubtarget<AArch64Subtarget>().isTargetDarwin()) { |
| 285 | if (SCS) |
| 286 | report_fatal_error(reason: "ShadowCallStack attribute not supported on Darwin." ); |
| 287 | return getDarwinCallPreservedMask(MF, CC); |
| 288 | } |
| 289 | |
| 290 | if (CC == CallingConv::AArch64_VectorCall) |
| 291 | return SCS ? CSR_AArch64_AAVPCS_SCS_RegMask : CSR_AArch64_AAVPCS_RegMask; |
| 292 | if (CC == CallingConv::AArch64_SVE_VectorCall) |
| 293 | return SCS ? CSR_AArch64_SVE_AAPCS_SCS_RegMask |
| 294 | : CSR_AArch64_SVE_AAPCS_RegMask; |
| 295 | if (CC == CallingConv::AArch64_SME_ABI_Support_Routines_PreserveMost_From_X0) |
| 296 | return CSR_AArch64_SME_ABI_Support_Routines_PreserveMost_From_X0_RegMask; |
| 297 | if (CC == CallingConv::AArch64_SME_ABI_Support_Routines_PreserveMost_From_X1) |
| 298 | return CSR_AArch64_SME_ABI_Support_Routines_PreserveMost_From_X1_RegMask; |
| 299 | if (CC == CallingConv::AArch64_SME_ABI_Support_Routines_PreserveMost_From_X2) |
| 300 | return CSR_AArch64_SME_ABI_Support_Routines_PreserveMost_From_X2_RegMask; |
| 301 | if (CC == CallingConv::CFGuard_Check) |
| 302 | return CSR_Win_AArch64_CFGuard_Check_RegMask; |
| 303 | if (MF.getSubtarget<AArch64Subtarget>().getTargetLowering() |
| 304 | ->supportSwiftError() && |
| 305 | MF.getFunction().getAttributes().hasAttrSomewhere(Kind: Attribute::SwiftError)) |
| 306 | return SCS ? CSR_AArch64_AAPCS_SwiftError_SCS_RegMask |
| 307 | : CSR_AArch64_AAPCS_SwiftError_RegMask; |
| 308 | if (CC == CallingConv::SwiftTail) { |
| 309 | if (SCS) |
| 310 | report_fatal_error(reason: "ShadowCallStack attribute not supported with swifttail" ); |
| 311 | return CSR_AArch64_AAPCS_SwiftTail_RegMask; |
| 312 | } |
| 313 | if (CC == CallingConv::PreserveMost) |
| 314 | return SCS ? CSR_AArch64_RT_MostRegs_SCS_RegMask |
| 315 | : CSR_AArch64_RT_MostRegs_RegMask; |
| 316 | if (CC == CallingConv::PreserveAll) |
| 317 | return SCS ? CSR_AArch64_RT_AllRegs_SCS_RegMask |
| 318 | : CSR_AArch64_RT_AllRegs_RegMask; |
| 319 | |
| 320 | return SCS ? CSR_AArch64_AAPCS_SCS_RegMask : CSR_AArch64_AAPCS_RegMask; |
| 321 | } |
| 322 | |
| 323 | const uint32_t *AArch64RegisterInfo::getCustomEHPadPreservedMask( |
| 324 | const MachineFunction &MF) const { |
| 325 | if (MF.getSubtarget<AArch64Subtarget>().isTargetLinux()) |
| 326 | return CSR_AArch64_AAPCS_RegMask; |
| 327 | |
| 328 | return nullptr; |
| 329 | } |
| 330 | |
| 331 | const uint32_t *AArch64RegisterInfo::getTLSCallPreservedMask() const { |
| 332 | if (TT.isOSDarwin()) |
| 333 | return CSR_Darwin_AArch64_TLS_RegMask; |
| 334 | |
| 335 | assert(TT.isOSBinFormatELF() && "Invalid target" ); |
| 336 | return CSR_AArch64_TLS_ELF_RegMask; |
| 337 | } |
| 338 | |
| 339 | void AArch64RegisterInfo::UpdateCustomCallPreservedMask(MachineFunction &MF, |
| 340 | const uint32_t **Mask) const { |
| 341 | uint32_t *UpdatedMask = MF.allocateRegMask(); |
| 342 | unsigned RegMaskSize = MachineOperand::getRegMaskSize(NumRegs: getNumRegs()); |
| 343 | memcpy(dest: UpdatedMask, src: *Mask, n: sizeof(UpdatedMask[0]) * RegMaskSize); |
| 344 | |
| 345 | for (size_t i = 0; i < AArch64::GPR64commonRegClass.getNumRegs(); ++i) { |
| 346 | if (MF.getSubtarget<AArch64Subtarget>().isXRegCustomCalleeSaved(i)) { |
| 347 | for (MCPhysReg SubReg : |
| 348 | subregs_inclusive(Reg: AArch64::GPR64commonRegClass.getRegister(i))) { |
| 349 | // See TargetRegisterInfo::getCallPreservedMask for how to interpret the |
| 350 | // register mask. |
| 351 | UpdatedMask[SubReg / 32] |= 1u << (SubReg % 32); |
| 352 | } |
| 353 | } |
| 354 | } |
| 355 | *Mask = UpdatedMask; |
| 356 | } |
| 357 | |
| 358 | const uint32_t *AArch64RegisterInfo::getSMStartStopCallPreservedMask() const { |
| 359 | return CSR_AArch64_SMStartStop_RegMask; |
| 360 | } |
| 361 | |
| 362 | const uint32_t * |
| 363 | AArch64RegisterInfo::SMEABISupportRoutinesCallPreservedMaskFromX0() const { |
| 364 | return CSR_AArch64_SME_ABI_Support_Routines_PreserveMost_From_X0_RegMask; |
| 365 | } |
| 366 | |
| 367 | const uint32_t *AArch64RegisterInfo::getNoPreservedMask() const { |
| 368 | return CSR_AArch64_NoRegs_RegMask; |
| 369 | } |
| 370 | |
| 371 | const uint32_t * |
| 372 | AArch64RegisterInfo::getThisReturnPreservedMask(const MachineFunction &MF, |
| 373 | CallingConv::ID CC) const { |
| 374 | // This should return a register mask that is the same as that returned by |
| 375 | // getCallPreservedMask but that additionally preserves the register used for |
| 376 | // the first i64 argument (which must also be the register used to return a |
| 377 | // single i64 return value) |
| 378 | // |
| 379 | // In case that the calling convention does not use the same register for |
| 380 | // both, the function should return NULL (does not currently apply) |
| 381 | assert(CC != CallingConv::GHC && "should not be GHC calling convention." ); |
| 382 | if (MF.getSubtarget<AArch64Subtarget>().isTargetDarwin()) |
| 383 | return CSR_Darwin_AArch64_AAPCS_ThisReturn_RegMask; |
| 384 | return CSR_AArch64_AAPCS_ThisReturn_RegMask; |
| 385 | } |
| 386 | |
| 387 | const uint32_t *AArch64RegisterInfo::getWindowsStackProbePreservedMask() const { |
| 388 | return CSR_AArch64_StackProbe_Windows_RegMask; |
| 389 | } |
| 390 | |
| 391 | std::optional<std::string> |
| 392 | AArch64RegisterInfo::explainReservedReg(const MachineFunction &MF, |
| 393 | MCRegister PhysReg) const { |
| 394 | if (hasBasePointer(MF) && MCRegisterInfo::regsOverlap(RegA: PhysReg, RegB: AArch64::X19)) |
| 395 | return std::string("X19 is used as the frame base pointer register." ); |
| 396 | |
| 397 | if (MF.getSubtarget<AArch64Subtarget>().isWindowsArm64EC()) { |
| 398 | bool warn = false; |
| 399 | if (MCRegisterInfo::regsOverlap(RegA: PhysReg, RegB: AArch64::X13) || |
| 400 | MCRegisterInfo::regsOverlap(RegA: PhysReg, RegB: AArch64::X14) || |
| 401 | MCRegisterInfo::regsOverlap(RegA: PhysReg, RegB: AArch64::X23) || |
| 402 | MCRegisterInfo::regsOverlap(RegA: PhysReg, RegB: AArch64::X24) || |
| 403 | MCRegisterInfo::regsOverlap(RegA: PhysReg, RegB: AArch64::X28)) |
| 404 | warn = true; |
| 405 | |
| 406 | for (unsigned i = AArch64::B16; i <= AArch64::B31; ++i) |
| 407 | if (MCRegisterInfo::regsOverlap(RegA: PhysReg, RegB: i)) |
| 408 | warn = true; |
| 409 | |
| 410 | if (warn) |
| 411 | return std::string(AArch64InstPrinter::getRegisterName(Reg: PhysReg)) + |
| 412 | " is clobbered by asynchronous signals when using Arm64EC." ; |
| 413 | } |
| 414 | |
| 415 | return {}; |
| 416 | } |
| 417 | |
| 418 | BitVector |
| 419 | AArch64RegisterInfo::getStrictlyReservedRegs(const MachineFunction &MF) const { |
| 420 | const AArch64FrameLowering *TFI = getFrameLowering(MF); |
| 421 | |
| 422 | // FIXME: avoid re-calculating this every time. |
| 423 | BitVector Reserved(getNumRegs()); |
| 424 | markSuperRegs(RegisterSet&: Reserved, Reg: AArch64::WSP); |
| 425 | markSuperRegs(RegisterSet&: Reserved, Reg: AArch64::WZR); |
| 426 | |
| 427 | if (TFI->isFPReserved(MF)) |
| 428 | markSuperRegs(RegisterSet&: Reserved, Reg: AArch64::W29); |
| 429 | |
| 430 | if (MF.getSubtarget<AArch64Subtarget>().isWindowsArm64EC()) { |
| 431 | // x13, x14, x23, x24, x28, and v16-v31 are clobbered by asynchronous |
| 432 | // signals, so we can't ever use them. |
| 433 | markSuperRegs(RegisterSet&: Reserved, Reg: AArch64::W13); |
| 434 | markSuperRegs(RegisterSet&: Reserved, Reg: AArch64::W14); |
| 435 | markSuperRegs(RegisterSet&: Reserved, Reg: AArch64::W23); |
| 436 | markSuperRegs(RegisterSet&: Reserved, Reg: AArch64::W24); |
| 437 | markSuperRegs(RegisterSet&: Reserved, Reg: AArch64::W28); |
| 438 | for (unsigned i = AArch64::B16; i <= AArch64::B31; ++i) |
| 439 | markSuperRegs(RegisterSet&: Reserved, Reg: i); |
| 440 | } |
| 441 | |
| 442 | if (MF.getSubtarget<AArch64Subtarget>().isLFI()) { |
| 443 | markSuperRegs(RegisterSet&: Reserved, Reg: AArch64::W28); |
| 444 | markSuperRegs(RegisterSet&: Reserved, Reg: AArch64::W27); |
| 445 | markSuperRegs(RegisterSet&: Reserved, Reg: AArch64::W26); |
| 446 | markSuperRegs(RegisterSet&: Reserved, Reg: AArch64::W25); |
| 447 | if (!MF.getProperties().hasNoVRegs()) { |
| 448 | markSuperRegs(RegisterSet&: Reserved, Reg: AArch64::LR); |
| 449 | markSuperRegs(RegisterSet&: Reserved, Reg: AArch64::W30); |
| 450 | } |
| 451 | } |
| 452 | |
| 453 | for (size_t i = 0; i < AArch64::GPR32commonRegClass.getNumRegs(); ++i) { |
| 454 | if (MF.getSubtarget<AArch64Subtarget>().isXRegisterReserved(i)) |
| 455 | markSuperRegs(RegisterSet&: Reserved, Reg: AArch64::GPR32commonRegClass.getRegister(i)); |
| 456 | } |
| 457 | |
| 458 | if (hasBasePointer(MF)) |
| 459 | markSuperRegs(RegisterSet&: Reserved, Reg: AArch64::W19); |
| 460 | |
| 461 | // SLH uses register W16/X16 as the taint register. |
| 462 | if (MF.getFunction().hasFnAttribute(Kind: Attribute::SpeculativeLoadHardening)) |
| 463 | markSuperRegs(RegisterSet&: Reserved, Reg: AArch64::W16); |
| 464 | |
| 465 | // FFR is modelled as global state that cannot be allocated. |
| 466 | if (MF.getSubtarget<AArch64Subtarget>().hasSVE()) |
| 467 | Reserved.set(AArch64::FFR); |
| 468 | |
| 469 | // SME tiles are not allocatable. |
| 470 | if (MF.getSubtarget<AArch64Subtarget>().hasSME()) { |
| 471 | for (MCPhysReg SubReg : subregs_inclusive(Reg: AArch64::ZA)) |
| 472 | Reserved.set(SubReg); |
| 473 | } |
| 474 | |
| 475 | // VG cannot be allocated |
| 476 | Reserved.set(AArch64::VG); |
| 477 | |
| 478 | if (MF.getSubtarget<AArch64Subtarget>().hasSME2()) { |
| 479 | for (MCSubRegIterator SubReg(AArch64::ZT0, this, /*self=*/true); |
| 480 | SubReg.isValid(); ++SubReg) |
| 481 | Reserved.set(*SubReg); |
| 482 | } |
| 483 | |
| 484 | markSuperRegs(RegisterSet&: Reserved, Reg: AArch64::FPCR); |
| 485 | markSuperRegs(RegisterSet&: Reserved, Reg: AArch64::FPMR); |
| 486 | markSuperRegs(RegisterSet&: Reserved, Reg: AArch64::FPSR); |
| 487 | |
| 488 | if (MF.getFunction().getCallingConv() == CallingConv::GRAAL) { |
| 489 | markSuperRegs(RegisterSet&: Reserved, Reg: AArch64::X27); |
| 490 | markSuperRegs(RegisterSet&: Reserved, Reg: AArch64::X28); |
| 491 | markSuperRegs(RegisterSet&: Reserved, Reg: AArch64::W27); |
| 492 | markSuperRegs(RegisterSet&: Reserved, Reg: AArch64::W28); |
| 493 | } |
| 494 | |
| 495 | assert(checkAllSuperRegsMarked(Reserved)); |
| 496 | |
| 497 | // Add _HI registers after checkAllSuperRegsMarked as this check otherwise |
| 498 | // becomes considerably more expensive. |
| 499 | Reserved.set(AArch64::WSP_HI); |
| 500 | Reserved.set(AArch64::WZR_HI); |
| 501 | static_assert(AArch64::W30_HI - AArch64::W0_HI == 30, |
| 502 | "Unexpected order of registers" ); |
| 503 | Reserved.set(I: AArch64::W0_HI, E: AArch64::W30_HI); |
| 504 | static_assert(AArch64::B31_HI - AArch64::B0_HI == 31, |
| 505 | "Unexpected order of registers" ); |
| 506 | Reserved.set(I: AArch64::B0_HI, E: AArch64::B31_HI); |
| 507 | static_assert(AArch64::H31_HI - AArch64::H0_HI == 31, |
| 508 | "Unexpected order of registers" ); |
| 509 | Reserved.set(I: AArch64::H0_HI, E: AArch64::H31_HI); |
| 510 | static_assert(AArch64::S31_HI - AArch64::S0_HI == 31, |
| 511 | "Unexpected order of registers" ); |
| 512 | Reserved.set(I: AArch64::S0_HI, E: AArch64::S31_HI); |
| 513 | static_assert(AArch64::D31_HI - AArch64::D0_HI == 31, |
| 514 | "Unexpected order of registers" ); |
| 515 | Reserved.set(I: AArch64::D0_HI, E: AArch64::D31_HI); |
| 516 | static_assert(AArch64::Q31_HI - AArch64::Q0_HI == 31, |
| 517 | "Unexpected order of registers" ); |
| 518 | Reserved.set(I: AArch64::Q0_HI, E: AArch64::Q31_HI); |
| 519 | |
| 520 | return Reserved; |
| 521 | } |
| 522 | |
| 523 | BitVector |
| 524 | AArch64RegisterInfo::getUserReservedRegs(const MachineFunction &MF) const { |
| 525 | BitVector Reserved(getNumRegs()); |
| 526 | for (size_t i = 0; i < AArch64::GPR32commonRegClass.getNumRegs(); ++i) { |
| 527 | // ReserveXRegister is set for registers manually reserved |
| 528 | // through +reserve-x#i. |
| 529 | if (MF.getSubtarget<AArch64Subtarget>().isXRegisterReserved(i)) |
| 530 | markSuperRegs(RegisterSet&: Reserved, Reg: AArch64::GPR32commonRegClass.getRegister(i)); |
| 531 | } |
| 532 | return Reserved; |
| 533 | } |
| 534 | |
| 535 | BitVector |
| 536 | AArch64RegisterInfo::getReservedRegs(const MachineFunction &MF) const { |
| 537 | BitVector Reserved(getNumRegs()); |
| 538 | for (size_t i = 0; i < AArch64::GPR32commonRegClass.getNumRegs(); ++i) { |
| 539 | if (MF.getSubtarget<AArch64Subtarget>().isXRegisterReservedForRA(i)) |
| 540 | markSuperRegs(RegisterSet&: Reserved, Reg: AArch64::GPR32commonRegClass.getRegister(i)); |
| 541 | } |
| 542 | |
| 543 | if (MF.getSubtarget<AArch64Subtarget>().isLRReservedForRA()) { |
| 544 | // In order to prevent the register allocator from using LR, we need to |
| 545 | // mark it as reserved. However we don't want to keep it reserved throughout |
| 546 | // the pipeline since it prevents other infrastructure from reasoning about |
| 547 | // it's liveness. We use the NoVRegs property instead of IsSSA because |
| 548 | // IsSSA is removed before VirtRegRewriter runs. |
| 549 | if (!MF.getProperties().hasNoVRegs()) |
| 550 | // Reserve LR (X30) by marking from its subregister W30 because otherwise |
| 551 | // the register allocator could clobber the subregister. |
| 552 | markSuperRegs(RegisterSet&: Reserved, Reg: AArch64::W30); |
| 553 | } |
| 554 | |
| 555 | assert(checkAllSuperRegsMarked(Reserved)); |
| 556 | |
| 557 | // Handle strictlyReservedRegs separately to avoid re-evaluating the assert, |
| 558 | // which becomes considerably expensive when considering the _HI registers. |
| 559 | Reserved |= getStrictlyReservedRegs(MF); |
| 560 | |
| 561 | return Reserved; |
| 562 | } |
| 563 | |
| 564 | bool AArch64RegisterInfo::isReservedReg(const MachineFunction &MF, |
| 565 | MCRegister Reg) const { |
| 566 | return getReservedRegs(MF)[Reg]; |
| 567 | } |
| 568 | |
| 569 | bool AArch64RegisterInfo::isUserReservedReg(const MachineFunction &MF, |
| 570 | MCRegister Reg) const { |
| 571 | return getUserReservedRegs(MF)[Reg]; |
| 572 | } |
| 573 | |
| 574 | bool AArch64RegisterInfo::isStrictlyReservedReg(const MachineFunction &MF, |
| 575 | MCRegister Reg) const { |
| 576 | return getStrictlyReservedRegs(MF)[Reg]; |
| 577 | } |
| 578 | |
| 579 | bool AArch64RegisterInfo::isAnyArgRegReserved(const MachineFunction &MF) const { |
| 580 | return llvm::any_of(Range: *AArch64::GPR64argRegClass.MC, P: [this, &MF](MCPhysReg r) { |
| 581 | return isStrictlyReservedReg(MF, Reg: r); |
| 582 | }); |
| 583 | } |
| 584 | |
| 585 | void AArch64RegisterInfo::emitReservedArgRegCallError( |
| 586 | const MachineFunction &MF) const { |
| 587 | const Function &F = MF.getFunction(); |
| 588 | F.getContext().diagnose(DI: DiagnosticInfoUnsupported{F, ("AArch64 doesn't support" |
| 589 | " function calls if any of the argument registers is reserved." )}); |
| 590 | } |
| 591 | |
| 592 | bool AArch64RegisterInfo::isAsmClobberable(const MachineFunction &MF, |
| 593 | MCRegister PhysReg) const { |
| 594 | // SLH uses register X16 as the taint register but it will fallback to a different |
| 595 | // method if the user clobbers it. So X16 is not reserved for inline asm but is |
| 596 | // for normal codegen. |
| 597 | if (MF.getFunction().hasFnAttribute(Kind: Attribute::SpeculativeLoadHardening) && |
| 598 | MCRegisterInfo::regsOverlap(RegA: PhysReg, RegB: AArch64::X16)) |
| 599 | return true; |
| 600 | |
| 601 | // ZA/ZT0 registers are reserved but may be permitted in the clobber list. |
| 602 | if (PhysReg == AArch64::ZA || PhysReg == AArch64::ZT0) |
| 603 | return true; |
| 604 | |
| 605 | return !isReservedReg(MF, Reg: PhysReg); |
| 606 | } |
| 607 | |
| 608 | const TargetRegisterClass * |
| 609 | AArch64RegisterInfo::getPointerRegClass(unsigned Kind) const { |
| 610 | return &AArch64::GPR64spRegClass; |
| 611 | } |
| 612 | |
| 613 | const TargetRegisterClass * |
| 614 | AArch64RegisterInfo::getCrossCopyRegClass(const TargetRegisterClass *RC) const { |
| 615 | if (RC == &AArch64::CCRRegClass) |
| 616 | return &AArch64::GPR64RegClass; // Only MSR & MRS copy NZCV. |
| 617 | return RC; |
| 618 | } |
| 619 | |
| 620 | MCRegister AArch64RegisterInfo::getBaseRegister() const { return AArch64::X19; } |
| 621 | |
| 622 | bool AArch64RegisterInfo::hasBasePointer(const MachineFunction &MF) const { |
| 623 | const MachineFrameInfo &MFI = MF.getFrameInfo(); |
| 624 | |
| 625 | // In the presence of variable sized objects or funclets, if the fixed stack |
| 626 | // size is large enough that referencing from the FP won't result in things |
| 627 | // being in range relatively often, we can use a base pointer to allow access |
| 628 | // from the other direction like the SP normally works. |
| 629 | // |
| 630 | // Furthermore, if both variable sized objects are present, and the |
| 631 | // stack needs to be dynamically re-aligned, the base pointer is the only |
| 632 | // reliable way to reference the locals. |
| 633 | if (MFI.hasVarSizedObjects() || MF.hasEHFunclets()) { |
| 634 | if (hasStackRealignment(MF)) |
| 635 | return true; |
| 636 | |
| 637 | auto &ST = MF.getSubtarget<AArch64Subtarget>(); |
| 638 | const AArch64FunctionInfo *AFI = MF.getInfo<AArch64FunctionInfo>(); |
| 639 | if (ST.hasSVE() || ST.isStreaming()) { |
| 640 | // Frames that have variable sized objects and scalable SVE objects, |
| 641 | // should always use a basepointer. |
| 642 | if (!AFI->hasCalculatedStackSizeSVE() || AFI->hasSVEStackSize()) |
| 643 | return true; |
| 644 | } |
| 645 | |
| 646 | // Frames with hazard padding can have a large offset between the frame |
| 647 | // pointer and GPR locals, which includes the emergency spill slot. If the |
| 648 | // emergency spill slot is not within range of the load/store instructions |
| 649 | // (which have a signed 9-bit range), we will fail to compile if it is used. |
| 650 | // Since hasBasePointer() is called before we know if we have hazard padding |
| 651 | // or an emergency spill slot we need to enable the basepointer |
| 652 | // conservatively. |
| 653 | if (ST.getStreamingHazardSize() && |
| 654 | !AFI->getSMEFnAttrs().hasNonStreamingInterfaceAndBody()) { |
| 655 | return true; |
| 656 | } |
| 657 | |
| 658 | // Conservatively estimate whether the negative offset from the frame |
| 659 | // pointer will be sufficient to reach. If a function has a smallish |
| 660 | // frame, it's less likely to have lots of spills and callee saved |
| 661 | // space, so it's all more likely to be within range of the frame pointer. |
| 662 | // If it's wrong, we'll materialize the constant and still get to the |
| 663 | // object; it's just suboptimal. Negative offsets use the unscaled |
| 664 | // load/store instructions, which have a 9-bit signed immediate. |
| 665 | return MFI.getLocalFrameSize() >= 256; |
| 666 | } |
| 667 | |
| 668 | return false; |
| 669 | } |
| 670 | |
| 671 | bool AArch64RegisterInfo::isArgumentRegister(const MachineFunction &MF, |
| 672 | MCRegister Reg) const { |
| 673 | CallingConv::ID CC = MF.getFunction().getCallingConv(); |
| 674 | const AArch64Subtarget &STI = MF.getSubtarget<AArch64Subtarget>(); |
| 675 | bool IsVarArg = STI.isCallingConvWin64(CC: MF.getFunction().getCallingConv(), |
| 676 | IsVarArg: MF.getFunction().isVarArg()); |
| 677 | |
| 678 | auto HasReg = [](ArrayRef<MCRegister> RegList, MCRegister Reg) { |
| 679 | return llvm::is_contained(Range&: RegList, Element: Reg); |
| 680 | }; |
| 681 | |
| 682 | switch (CC) { |
| 683 | default: |
| 684 | report_fatal_error(reason: "Unsupported calling convention." ); |
| 685 | case CallingConv::GHC: |
| 686 | return HasReg(CC_AArch64_GHC_ArgRegs, Reg); |
| 687 | case CallingConv::PreserveNone: |
| 688 | if (!MF.getFunction().isVarArg()) |
| 689 | return HasReg(CC_AArch64_Preserve_None_ArgRegs, Reg); |
| 690 | [[fallthrough]]; |
| 691 | case CallingConv::C: |
| 692 | case CallingConv::Fast: |
| 693 | case CallingConv::PreserveMost: |
| 694 | case CallingConv::PreserveAll: |
| 695 | case CallingConv::CXX_FAST_TLS: |
| 696 | case CallingConv::Swift: |
| 697 | case CallingConv::SwiftTail: |
| 698 | case CallingConv::Tail: |
| 699 | if (STI.isTargetWindows()) { |
| 700 | if (IsVarArg) |
| 701 | return HasReg(CC_AArch64_Win64_VarArg_ArgRegs, Reg); |
| 702 | switch (CC) { |
| 703 | default: |
| 704 | return HasReg(CC_AArch64_Win64PCS_ArgRegs, Reg); |
| 705 | case CallingConv::Swift: |
| 706 | case CallingConv::SwiftTail: |
| 707 | return HasReg(CC_AArch64_Win64PCS_Swift_ArgRegs, Reg) || |
| 708 | HasReg(CC_AArch64_Win64PCS_ArgRegs, Reg); |
| 709 | } |
| 710 | } |
| 711 | if (!STI.isTargetDarwin()) { |
| 712 | switch (CC) { |
| 713 | default: |
| 714 | return HasReg(CC_AArch64_AAPCS_ArgRegs, Reg); |
| 715 | case CallingConv::Swift: |
| 716 | case CallingConv::SwiftTail: |
| 717 | return HasReg(CC_AArch64_AAPCS_ArgRegs, Reg) || |
| 718 | HasReg(CC_AArch64_AAPCS_Swift_ArgRegs, Reg); |
| 719 | } |
| 720 | } |
| 721 | if (!IsVarArg) { |
| 722 | switch (CC) { |
| 723 | default: |
| 724 | return HasReg(CC_AArch64_DarwinPCS_ArgRegs, Reg); |
| 725 | case CallingConv::Swift: |
| 726 | case CallingConv::SwiftTail: |
| 727 | return HasReg(CC_AArch64_DarwinPCS_ArgRegs, Reg) || |
| 728 | HasReg(CC_AArch64_DarwinPCS_Swift_ArgRegs, Reg); |
| 729 | } |
| 730 | } |
| 731 | if (STI.isTargetILP32()) |
| 732 | return HasReg(CC_AArch64_DarwinPCS_ILP32_VarArg_ArgRegs, Reg); |
| 733 | return HasReg(CC_AArch64_DarwinPCS_VarArg_ArgRegs, Reg); |
| 734 | case CallingConv::Win64: |
| 735 | if (IsVarArg) |
| 736 | HasReg(CC_AArch64_Win64_VarArg_ArgRegs, Reg); |
| 737 | return HasReg(CC_AArch64_Win64PCS_ArgRegs, Reg); |
| 738 | case CallingConv::CFGuard_Check: |
| 739 | return HasReg(CC_AArch64_Win64_CFGuard_Check_ArgRegs, Reg); |
| 740 | case CallingConv::AArch64_VectorCall: |
| 741 | case CallingConv::AArch64_SVE_VectorCall: |
| 742 | case CallingConv::AArch64_SME_ABI_Support_Routines_PreserveMost_From_X0: |
| 743 | case CallingConv::AArch64_SME_ABI_Support_Routines_PreserveMost_From_X1: |
| 744 | case CallingConv::AArch64_SME_ABI_Support_Routines_PreserveMost_From_X2: |
| 745 | if (STI.isTargetWindows()) |
| 746 | return HasReg(CC_AArch64_Win64PCS_ArgRegs, Reg); |
| 747 | return HasReg(CC_AArch64_AAPCS_ArgRegs, Reg); |
| 748 | } |
| 749 | } |
| 750 | |
| 751 | Register |
| 752 | AArch64RegisterInfo::getFrameRegister(const MachineFunction &MF) const { |
| 753 | const AArch64FrameLowering *TFI = getFrameLowering(MF); |
| 754 | return TFI->hasFP(MF) ? AArch64::FP : AArch64::SP; |
| 755 | } |
| 756 | |
| 757 | bool AArch64RegisterInfo::requiresRegisterScavenging( |
| 758 | const MachineFunction &MF) const { |
| 759 | return true; |
| 760 | } |
| 761 | |
| 762 | bool AArch64RegisterInfo::requiresVirtualBaseRegisters( |
| 763 | const MachineFunction &MF) const { |
| 764 | return true; |
| 765 | } |
| 766 | |
| 767 | bool |
| 768 | AArch64RegisterInfo::useFPForScavengingIndex(const MachineFunction &MF) const { |
| 769 | // This function indicates whether the emergency spillslot should be placed |
| 770 | // close to the beginning of the stackframe (closer to FP) or the end |
| 771 | // (closer to SP). |
| 772 | // |
| 773 | // The beginning works most reliably if we have a frame pointer. |
| 774 | // In the presence of any non-constant space between FP and locals, |
| 775 | // (e.g. in case of stack realignment or a scalable SVE area), it is |
| 776 | // better to use SP or BP. |
| 777 | const AArch64FrameLowering &TFI = *getFrameLowering(MF); |
| 778 | const AArch64FunctionInfo *AFI = MF.getInfo<AArch64FunctionInfo>(); |
| 779 | assert((!MF.getSubtarget<AArch64Subtarget>().hasSVE() || |
| 780 | AFI->hasCalculatedStackSizeSVE()) && |
| 781 | "Expected SVE area to be calculated by this point" ); |
| 782 | return TFI.hasFP(MF) && !hasStackRealignment(MF) && !AFI->hasSVEStackSize() && |
| 783 | !AFI->hasStackHazardSlotIndex(); |
| 784 | } |
| 785 | |
| 786 | bool AArch64RegisterInfo::requiresFrameIndexScavenging( |
| 787 | const MachineFunction &MF) const { |
| 788 | return true; |
| 789 | } |
| 790 | |
| 791 | bool |
| 792 | AArch64RegisterInfo::cannotEliminateFrame(const MachineFunction &MF) const { |
| 793 | const MachineFrameInfo &MFI = MF.getFrameInfo(); |
| 794 | if (MF.getTarget().Options.DisableFramePointerElim(MF) && MFI.adjustsStack()) |
| 795 | return true; |
| 796 | return MFI.hasVarSizedObjects() || MFI.isFrameAddressTaken(); |
| 797 | } |
| 798 | |
| 799 | /// needsFrameBaseReg - Returns true if the instruction's frame index |
| 800 | /// reference would be better served by a base register other than FP |
| 801 | /// or SP. Used by LocalStackFrameAllocation to determine which frame index |
| 802 | /// references it should create new base registers for. |
| 803 | bool AArch64RegisterInfo::needsFrameBaseReg(MachineInstr *MI, |
| 804 | int64_t Offset) const { |
| 805 | for (unsigned i = 0; !MI->getOperand(i).isFI(); ++i) |
| 806 | assert(i < MI->getNumOperands() && |
| 807 | "Instr doesn't have FrameIndex operand!" ); |
| 808 | |
| 809 | // It's the load/store FI references that cause issues, as it can be difficult |
| 810 | // to materialize the offset if it won't fit in the literal field. Estimate |
| 811 | // based on the size of the local frame and some conservative assumptions |
| 812 | // about the rest of the stack frame (note, this is pre-regalloc, so |
| 813 | // we don't know everything for certain yet) whether this offset is likely |
| 814 | // to be out of range of the immediate. Return true if so. |
| 815 | |
| 816 | // We only generate virtual base registers for loads and stores, so |
| 817 | // return false for everything else. |
| 818 | if (!MI->mayLoad() && !MI->mayStore()) |
| 819 | return false; |
| 820 | |
| 821 | // Without a virtual base register, if the function has variable sized |
| 822 | // objects, all fixed-size local references will be via the frame pointer, |
| 823 | // Approximate the offset and see if it's legal for the instruction. |
| 824 | // Note that the incoming offset is based on the SP value at function entry, |
| 825 | // so it'll be negative. |
| 826 | MachineFunction &MF = *MI->getParent()->getParent(); |
| 827 | const AArch64FrameLowering *TFI = getFrameLowering(MF); |
| 828 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
| 829 | |
| 830 | // Estimate an offset from the frame pointer. |
| 831 | // Conservatively assume all GPR callee-saved registers get pushed. |
| 832 | // FP, LR, X19-X28, D8-D15. 64-bits each. |
| 833 | int64_t FPOffset = Offset - 16 * 20; |
| 834 | // Estimate an offset from the stack pointer. |
| 835 | // The incoming offset is relating to the SP at the start of the function, |
| 836 | // but when we access the local it'll be relative to the SP after local |
| 837 | // allocation, so adjust our SP-relative offset by that allocation size. |
| 838 | Offset += MFI.getLocalFrameSize(); |
| 839 | // Assume that we'll have at least some spill slots allocated. |
| 840 | // FIXME: This is a total SWAG number. We should run some statistics |
| 841 | // and pick a real one. |
| 842 | Offset += 128; // 128 bytes of spill slots |
| 843 | |
| 844 | // If there is a frame pointer, try using it. |
| 845 | // The FP is only available if there is no dynamic realignment. We |
| 846 | // don't know for sure yet whether we'll need that, so we guess based |
| 847 | // on whether there are any local variables that would trigger it. |
| 848 | if (TFI->hasFP(MF) && isFrameOffsetLegal(MI, BaseReg: AArch64::FP, Offset: FPOffset)) |
| 849 | return false; |
| 850 | |
| 851 | // If we can reference via the stack pointer or base pointer, try that. |
| 852 | // FIXME: This (and the code that resolves the references) can be improved |
| 853 | // to only disallow SP relative references in the live range of |
| 854 | // the VLA(s). In practice, it's unclear how much difference that |
| 855 | // would make, but it may be worth doing. |
| 856 | if (isFrameOffsetLegal(MI, BaseReg: AArch64::SP, Offset)) |
| 857 | return false; |
| 858 | |
| 859 | // If even offset 0 is illegal, we don't want a virtual base register. |
| 860 | if (!isFrameOffsetLegal(MI, BaseReg: AArch64::SP, Offset: 0)) |
| 861 | return false; |
| 862 | |
| 863 | // The offset likely isn't legal; we want to allocate a virtual base register. |
| 864 | return true; |
| 865 | } |
| 866 | |
| 867 | bool AArch64RegisterInfo::isFrameOffsetLegal(const MachineInstr *MI, |
| 868 | Register BaseReg, |
| 869 | int64_t Offset) const { |
| 870 | assert(MI && "Unable to get the legal offset for nil instruction." ); |
| 871 | StackOffset SaveOffset = StackOffset::getFixed(Fixed: Offset); |
| 872 | return isAArch64FrameOffsetLegal(MI: *MI, Offset&: SaveOffset) & AArch64FrameOffsetIsLegal; |
| 873 | } |
| 874 | |
| 875 | /// Insert defining instruction(s) for BaseReg to be a pointer to FrameIdx |
| 876 | /// at the beginning of the basic block. |
| 877 | Register |
| 878 | AArch64RegisterInfo::materializeFrameBaseRegister(MachineBasicBlock *MBB, |
| 879 | int FrameIdx, |
| 880 | int64_t Offset) const { |
| 881 | MachineBasicBlock::iterator Ins = MBB->begin(); |
| 882 | DebugLoc DL; // Defaults to "unknown" |
| 883 | if (Ins != MBB->end()) |
| 884 | DL = Ins->getDebugLoc(); |
| 885 | const MachineFunction &MF = *MBB->getParent(); |
| 886 | const AArch64InstrInfo *TII = |
| 887 | MF.getSubtarget<AArch64Subtarget>().getInstrInfo(); |
| 888 | const MCInstrDesc &MCID = TII->get(Opcode: AArch64::ADDXri); |
| 889 | MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo(); |
| 890 | Register BaseReg = MRI.createVirtualRegister(RegClass: &AArch64::GPR64spRegClass); |
| 891 | MRI.constrainRegClass(Reg: BaseReg, RC: TII->getRegClass(MCID, OpNum: 0)); |
| 892 | unsigned Shifter = AArch64_AM::getShifterImm(ST: AArch64_AM::LSL, Imm: 0); |
| 893 | |
| 894 | BuildMI(BB&: *MBB, I: Ins, MIMD: DL, MCID, DestReg: BaseReg) |
| 895 | .addFrameIndex(Idx: FrameIdx) |
| 896 | .addImm(Val: Offset) |
| 897 | .addImm(Val: Shifter); |
| 898 | |
| 899 | return BaseReg; |
| 900 | } |
| 901 | |
| 902 | void AArch64RegisterInfo::resolveFrameIndex(MachineInstr &MI, Register BaseReg, |
| 903 | int64_t Offset) const { |
| 904 | // ARM doesn't need the general 64-bit offsets |
| 905 | StackOffset Off = StackOffset::getFixed(Fixed: Offset); |
| 906 | |
| 907 | unsigned i = 0; |
| 908 | while (!MI.getOperand(i).isFI()) { |
| 909 | ++i; |
| 910 | assert(i < MI.getNumOperands() && "Instr doesn't have FrameIndex operand!" ); |
| 911 | } |
| 912 | |
| 913 | const MachineFunction *MF = MI.getParent()->getParent(); |
| 914 | const AArch64InstrInfo *TII = |
| 915 | MF->getSubtarget<AArch64Subtarget>().getInstrInfo(); |
| 916 | bool Done = rewriteAArch64FrameIndex(MI, FrameRegIdx: i, FrameReg: BaseReg, Offset&: Off, TII); |
| 917 | assert(Done && "Unable to resolve frame index!" ); |
| 918 | (void)Done; |
| 919 | } |
| 920 | |
| 921 | // Create a scratch register for the frame index elimination in an instruction. |
| 922 | // This function has special handling of stack tagging loop pseudos, in which |
| 923 | // case it can also change the instruction opcode. |
| 924 | static Register |
| 925 | createScratchRegisterForInstruction(MachineInstr &MI, unsigned FIOperandNum, |
| 926 | const AArch64InstrInfo *TII) { |
| 927 | // ST*Gloop have a reserved scratch register in operand 1. Use it, and also |
| 928 | // replace the instruction with the writeback variant because it will now |
| 929 | // satisfy the operand constraints for it. |
| 930 | Register ScratchReg; |
| 931 | if (MI.getOpcode() == AArch64::STGloop || |
| 932 | MI.getOpcode() == AArch64::STZGloop) { |
| 933 | assert(FIOperandNum == 3 && |
| 934 | "Wrong frame index operand for STGloop/STZGloop" ); |
| 935 | unsigned Op = MI.getOpcode() == AArch64::STGloop ? AArch64::STGloop_wback |
| 936 | : AArch64::STZGloop_wback; |
| 937 | ScratchReg = MI.getOperand(i: 1).getReg(); |
| 938 | MI.getOperand(i: 3).ChangeToRegister(Reg: ScratchReg, isDef: false, isImp: false, isKill: true); |
| 939 | MI.setDesc(TII->get(Opcode: Op)); |
| 940 | MI.tieOperands(DefIdx: 1, UseIdx: 3); |
| 941 | } else { |
| 942 | ScratchReg = |
| 943 | MI.getMF()->getRegInfo().createVirtualRegister(RegClass: &AArch64::GPR64RegClass); |
| 944 | MI.getOperand(i: FIOperandNum) |
| 945 | .ChangeToRegister(Reg: ScratchReg, isDef: false, isImp: false, isKill: true); |
| 946 | } |
| 947 | return ScratchReg; |
| 948 | } |
| 949 | |
| 950 | void AArch64RegisterInfo::getOffsetOpcodes( |
| 951 | const StackOffset &Offset, SmallVectorImpl<uint64_t> &Ops) const { |
| 952 | // The smallest scalable element supported by scaled SVE addressing |
| 953 | // modes are predicates, which are 2 scalable bytes in size. So the scalable |
| 954 | // byte offset must always be a multiple of 2. |
| 955 | assert(Offset.getScalable() % 2 == 0 && "Invalid frame offset" ); |
| 956 | |
| 957 | // Add fixed-sized offset using existing DIExpression interface. |
| 958 | DIExpression::appendOffset(Ops, Offset: Offset.getFixed()); |
| 959 | |
| 960 | unsigned VG = getDwarfRegNum(Reg: AArch64::VG, isEH: true); |
| 961 | int64_t VGSized = Offset.getScalable() / 2; |
| 962 | if (VGSized > 0) { |
| 963 | Ops.push_back(Elt: dwarf::DW_OP_constu); |
| 964 | Ops.push_back(Elt: VGSized); |
| 965 | Ops.append(IL: {dwarf::DW_OP_bregx, VG, 0ULL}); |
| 966 | Ops.push_back(Elt: dwarf::DW_OP_mul); |
| 967 | Ops.push_back(Elt: dwarf::DW_OP_plus); |
| 968 | } else if (VGSized < 0) { |
| 969 | Ops.push_back(Elt: dwarf::DW_OP_constu); |
| 970 | Ops.push_back(Elt: -VGSized); |
| 971 | Ops.append(IL: {dwarf::DW_OP_bregx, VG, 0ULL}); |
| 972 | Ops.push_back(Elt: dwarf::DW_OP_mul); |
| 973 | Ops.push_back(Elt: dwarf::DW_OP_minus); |
| 974 | } |
| 975 | } |
| 976 | |
| 977 | bool AArch64RegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II, |
| 978 | int SPAdj, unsigned FIOperandNum, |
| 979 | RegScavenger *RS) const { |
| 980 | assert(SPAdj == 0 && "Unexpected" ); |
| 981 | |
| 982 | MachineInstr &MI = *II; |
| 983 | MachineBasicBlock &MBB = *MI.getParent(); |
| 984 | MachineFunction &MF = *MBB.getParent(); |
| 985 | const MachineFrameInfo &MFI = MF.getFrameInfo(); |
| 986 | const AArch64InstrInfo *TII = |
| 987 | MF.getSubtarget<AArch64Subtarget>().getInstrInfo(); |
| 988 | const AArch64FrameLowering *TFI = getFrameLowering(MF); |
| 989 | int FrameIndex = MI.getOperand(i: FIOperandNum).getIndex(); |
| 990 | bool Tagged = |
| 991 | MI.getOperand(i: FIOperandNum).getTargetFlags() & AArch64II::MO_TAGGED; |
| 992 | Register FrameReg; |
| 993 | |
| 994 | // Special handling of dbg_value, stackmap patchpoint statepoint instructions. |
| 995 | if (MI.getOpcode() == TargetOpcode::STACKMAP || |
| 996 | MI.getOpcode() == TargetOpcode::PATCHPOINT || |
| 997 | MI.getOpcode() == TargetOpcode::STATEPOINT) { |
| 998 | StackOffset Offset = |
| 999 | TFI->resolveFrameIndexReference(MF, FI: FrameIndex, FrameReg, |
| 1000 | /*PreferFP=*/true, |
| 1001 | /*ForSimm=*/false); |
| 1002 | Offset += StackOffset::getFixed(Fixed: MI.getOperand(i: FIOperandNum + 1).getImm()); |
| 1003 | MI.getOperand(i: FIOperandNum).ChangeToRegister(Reg: FrameReg, isDef: false /*isDef*/); |
| 1004 | MI.getOperand(i: FIOperandNum + 1).ChangeToImmediate(ImmVal: Offset.getFixed()); |
| 1005 | return false; |
| 1006 | } |
| 1007 | |
| 1008 | if (MI.getOpcode() == TargetOpcode::LOCAL_ESCAPE) { |
| 1009 | MachineOperand &FI = MI.getOperand(i: FIOperandNum); |
| 1010 | StackOffset Offset = TFI->getNonLocalFrameIndexReference(MF, FI: FrameIndex); |
| 1011 | assert(!Offset.getScalable() && |
| 1012 | "Frame offsets with a scalable component are not supported" ); |
| 1013 | FI.ChangeToImmediate(ImmVal: Offset.getFixed()); |
| 1014 | return false; |
| 1015 | } |
| 1016 | |
| 1017 | StackOffset Offset; |
| 1018 | if (MI.getOpcode() == AArch64::TAGPstack) { |
| 1019 | // TAGPstack must use the virtual frame register in its 3rd operand. |
| 1020 | const AArch64FunctionInfo *AFI = MF.getInfo<AArch64FunctionInfo>(); |
| 1021 | FrameReg = MI.getOperand(i: 3).getReg(); |
| 1022 | Offset = StackOffset::getFixed(Fixed: MFI.getObjectOffset(ObjectIdx: FrameIndex) + |
| 1023 | AFI->getTaggedBasePointerOffset()); |
| 1024 | } else if (Tagged) { |
| 1025 | StackOffset SPOffset = StackOffset::getFixed( |
| 1026 | Fixed: MFI.getObjectOffset(ObjectIdx: FrameIndex) + (int64_t)MFI.getStackSize()); |
| 1027 | if (MFI.hasVarSizedObjects() || |
| 1028 | isAArch64FrameOffsetLegal(MI, Offset&: SPOffset, OutUseUnscaledOp: nullptr, OutUnscaledOp: nullptr, EmittableOffset: nullptr) != |
| 1029 | (AArch64FrameOffsetCanUpdate | AArch64FrameOffsetIsLegal)) { |
| 1030 | // Can't update to SP + offset in place. Precalculate the tagged pointer |
| 1031 | // in a scratch register. |
| 1032 | Offset = TFI->resolveFrameIndexReference( |
| 1033 | MF, FI: FrameIndex, FrameReg, /*PreferFP=*/false, /*ForSimm=*/true); |
| 1034 | Register ScratchReg = |
| 1035 | MF.getRegInfo().createVirtualRegister(RegClass: &AArch64::GPR64RegClass); |
| 1036 | emitFrameOffset(MBB, MBBI: II, DL: MI.getDebugLoc(), DestReg: ScratchReg, SrcReg: FrameReg, Offset, |
| 1037 | TII); |
| 1038 | BuildMI(BB&: MBB, I&: MI, MIMD: MI.getDebugLoc(), MCID: TII->get(Opcode: AArch64::LDG), DestReg: ScratchReg) |
| 1039 | .addReg(RegNo: ScratchReg) |
| 1040 | .addReg(RegNo: ScratchReg) |
| 1041 | .addImm(Val: 0); |
| 1042 | MI.getOperand(i: FIOperandNum) |
| 1043 | .ChangeToRegister(Reg: ScratchReg, isDef: false, isImp: false, isKill: true); |
| 1044 | return false; |
| 1045 | } |
| 1046 | FrameReg = AArch64::SP; |
| 1047 | Offset = StackOffset::getFixed(Fixed: MFI.getObjectOffset(ObjectIdx: FrameIndex) + |
| 1048 | (int64_t)MFI.getStackSize()); |
| 1049 | } else { |
| 1050 | Offset = TFI->resolveFrameIndexReference( |
| 1051 | MF, FI: FrameIndex, FrameReg, /*PreferFP=*/false, /*ForSimm=*/true); |
| 1052 | } |
| 1053 | |
| 1054 | // Modify MI as necessary to handle as much of 'Offset' as possible |
| 1055 | if (rewriteAArch64FrameIndex(MI, FrameRegIdx: FIOperandNum, FrameReg, Offset, TII)) |
| 1056 | return true; |
| 1057 | |
| 1058 | assert((!RS || !RS->isScavengingFrameIndex(FrameIndex)) && |
| 1059 | "Emergency spill slot is out of reach" ); |
| 1060 | |
| 1061 | // If we get here, the immediate doesn't fit into the instruction. We folded |
| 1062 | // as much as possible above. Handle the rest, providing a register that is |
| 1063 | // SP+LargeImm. |
| 1064 | Register ScratchReg = |
| 1065 | createScratchRegisterForInstruction(MI, FIOperandNum, TII); |
| 1066 | emitFrameOffset(MBB, MBBI: II, DL: MI.getDebugLoc(), DestReg: ScratchReg, SrcReg: FrameReg, Offset, TII); |
| 1067 | return false; |
| 1068 | } |
| 1069 | |
| 1070 | unsigned AArch64RegisterInfo::getRegPressureLimit(const TargetRegisterClass *RC, |
| 1071 | MachineFunction &MF) const { |
| 1072 | const AArch64FrameLowering *TFI = getFrameLowering(MF); |
| 1073 | |
| 1074 | switch (RC->getID()) { |
| 1075 | default: |
| 1076 | return 0; |
| 1077 | case AArch64::GPR32RegClassID: |
| 1078 | case AArch64::GPR32spRegClassID: |
| 1079 | case AArch64::GPR32allRegClassID: |
| 1080 | case AArch64::GPR64spRegClassID: |
| 1081 | case AArch64::GPR64allRegClassID: |
| 1082 | case AArch64::GPR64RegClassID: |
| 1083 | case AArch64::GPR32commonRegClassID: |
| 1084 | case AArch64::GPR64commonRegClassID: |
| 1085 | return 32 - 1 // XZR/SP |
| 1086 | - (TFI->hasFP(MF) || TT.isOSDarwin()) // FP |
| 1087 | - MF.getSubtarget<AArch64Subtarget>().getNumXRegisterReserved() |
| 1088 | - hasBasePointer(MF); // X19 |
| 1089 | case AArch64::FPR8RegClassID: |
| 1090 | case AArch64::FPR16RegClassID: |
| 1091 | case AArch64::FPR32RegClassID: |
| 1092 | case AArch64::FPR64RegClassID: |
| 1093 | case AArch64::FPR128RegClassID: |
| 1094 | return 32; |
| 1095 | |
| 1096 | case AArch64::MatrixIndexGPR32_8_11RegClassID: |
| 1097 | case AArch64::MatrixIndexGPR32_12_15RegClassID: |
| 1098 | return 4; |
| 1099 | |
| 1100 | case AArch64::DDRegClassID: |
| 1101 | case AArch64::DDDRegClassID: |
| 1102 | case AArch64::DDDDRegClassID: |
| 1103 | case AArch64::QQRegClassID: |
| 1104 | case AArch64::QQQRegClassID: |
| 1105 | case AArch64::QQQQRegClassID: |
| 1106 | return 32; |
| 1107 | |
| 1108 | case AArch64::FPR128_loRegClassID: |
| 1109 | case AArch64::FPR64_loRegClassID: |
| 1110 | case AArch64::FPR16_loRegClassID: |
| 1111 | return 16; |
| 1112 | case AArch64::FPR128_0to7RegClassID: |
| 1113 | return 8; |
| 1114 | } |
| 1115 | } |
| 1116 | |
| 1117 | // We add regalloc hints for different cases: |
| 1118 | // * Choosing a better destination operand for predicated SVE instructions |
| 1119 | // where the inactive lanes are undef, by choosing a register that is not |
| 1120 | // unique to the other operands of the instruction. |
| 1121 | // |
| 1122 | // * Improve register allocation for SME multi-vector instructions where we can |
| 1123 | // benefit from the strided- and contiguous register multi-vector tuples. |
| 1124 | // |
| 1125 | // Here FORM_TRANSPOSED_REG_TUPLE nodes are created to improve register |
| 1126 | // allocation where a consecutive multi-vector tuple is constructed from the |
| 1127 | // same indices of multiple strided loads. This may still result in |
| 1128 | // unnecessary copies between the loads and the tuple. Here we try to return a |
| 1129 | // hint to assign the contiguous ZPRMulReg starting at the same register as |
| 1130 | // the first operand of the pseudo, which should be a subregister of the first |
| 1131 | // strided load. |
| 1132 | // |
| 1133 | // For example, if the first strided load has been assigned $z16_z20_z24_z28 |
| 1134 | // and the operands of the pseudo are each accessing subregister zsub2, we |
| 1135 | // should look through through Order to find a contiguous register which |
| 1136 | // begins with $z24 (i.e. $z24_z25_z26_z27). |
| 1137 | bool AArch64RegisterInfo::getRegAllocationHints( |
| 1138 | Register VirtReg, ArrayRef<MCPhysReg> Order, |
| 1139 | SmallVectorImpl<MCPhysReg> &Hints, const MachineFunction &MF, |
| 1140 | const VirtRegMap *VRM, const LiveRegMatrix *Matrix) const { |
| 1141 | auto &ST = MF.getSubtarget<AArch64Subtarget>(); |
| 1142 | const AArch64InstrInfo *TII = |
| 1143 | MF.getSubtarget<AArch64Subtarget>().getInstrInfo(); |
| 1144 | const MachineRegisterInfo &MRI = MF.getRegInfo(); |
| 1145 | |
| 1146 | // For predicated SVE instructions where the inactive lanes are undef, |
| 1147 | // pick a destination register that is not unique to avoid introducing |
| 1148 | // a movprfx. |
| 1149 | const TargetRegisterClass *RegRC = MRI.getRegClass(Reg: VirtReg); |
| 1150 | if (AArch64::ZPRRegClass.hasSubClassEq(RC: RegRC)) { |
| 1151 | bool ConsiderOnlyHints = TargetRegisterInfo::getRegAllocationHints( |
| 1152 | VirtReg, Order, Hints, MF, VRM); |
| 1153 | |
| 1154 | for (const MachineOperand &DefOp : MRI.def_operands(Reg: VirtReg)) { |
| 1155 | const MachineInstr &Def = *DefOp.getParent(); |
| 1156 | if (DefOp.isImplicit() || |
| 1157 | (TII->get(Opcode: Def.getOpcode()).TSFlags & AArch64::FalseLanesMask) != |
| 1158 | AArch64::FalseLanesUndef) |
| 1159 | continue; |
| 1160 | |
| 1161 | unsigned InstFlags = |
| 1162 | TII->get(Opcode: AArch64::getSVEPseudoMap(Opcode: Def.getOpcode())).TSFlags; |
| 1163 | |
| 1164 | for (MCPhysReg R : Order) { |
| 1165 | auto AddHintIfSuitable = [&](MCPhysReg R, |
| 1166 | const MachineOperand &MO) -> bool { |
| 1167 | // R is a suitable register hint if R can reuse one of the other |
| 1168 | // source operands. |
| 1169 | if (VRM->getPhys(virtReg: MO.getReg()) != R) |
| 1170 | return false; |
| 1171 | Hints.push_back(Elt: R); |
| 1172 | return true; |
| 1173 | }; |
| 1174 | |
| 1175 | switch (InstFlags & AArch64::DestructiveInstTypeMask) { |
| 1176 | default: |
| 1177 | break; |
| 1178 | case AArch64::DestructiveTernaryCommWithRev: |
| 1179 | AddHintIfSuitable(R, Def.getOperand(i: 2)) || |
| 1180 | AddHintIfSuitable(R, Def.getOperand(i: 3)) || |
| 1181 | AddHintIfSuitable(R, Def.getOperand(i: 4)); |
| 1182 | break; |
| 1183 | case AArch64::DestructiveBinaryComm: |
| 1184 | case AArch64::DestructiveBinaryCommWithRev: |
| 1185 | AddHintIfSuitable(R, Def.getOperand(i: 2)) || |
| 1186 | AddHintIfSuitable(R, Def.getOperand(i: 3)); |
| 1187 | break; |
| 1188 | case AArch64::DestructiveBinary: |
| 1189 | case AArch64::DestructiveBinaryImm: |
| 1190 | AddHintIfSuitable(R, Def.getOperand(i: 2)); |
| 1191 | break; |
| 1192 | case AArch64::DestructiveUnaryPassthru: |
| 1193 | AddHintIfSuitable(R, Def.getOperand(i: 3)); |
| 1194 | break; |
| 1195 | } |
| 1196 | } |
| 1197 | } |
| 1198 | |
| 1199 | if (Hints.size()) |
| 1200 | return ConsiderOnlyHints; |
| 1201 | } |
| 1202 | |
| 1203 | if (!ST.hasSME() || !ST.isStreaming()) |
| 1204 | return TargetRegisterInfo::getRegAllocationHints(VirtReg, Order, Hints, MF, |
| 1205 | VRM); |
| 1206 | |
| 1207 | // The SVE calling convention preserves registers Z8-Z23. As a result, there |
| 1208 | // are no ZPR2Strided or ZPR4Strided registers that do not overlap with the |
| 1209 | // callee-saved registers and so by default these will be pushed to the back |
| 1210 | // of the allocation order for the ZPRStridedOrContiguous classes. |
| 1211 | // If any of the instructions which define VirtReg are used by the |
| 1212 | // FORM_TRANSPOSED_REG_TUPLE pseudo, we want to favour reducing copy |
| 1213 | // instructions over reducing the number of clobbered callee-save registers, |
| 1214 | // so we add the strided registers as a hint. |
| 1215 | unsigned RegID = RegRC->getID(); |
| 1216 | if (RegID == AArch64::ZPR2StridedOrContiguousRegClassID || |
| 1217 | RegID == AArch64::ZPR4StridedOrContiguousRegClassID) { |
| 1218 | |
| 1219 | // Look through uses of the register for FORM_TRANSPOSED_REG_TUPLE. |
| 1220 | for (const MachineInstr &Use : MRI.use_nodbg_instructions(Reg: VirtReg)) { |
| 1221 | if (Use.getOpcode() != AArch64::FORM_TRANSPOSED_REG_TUPLE_X2_PSEUDO && |
| 1222 | Use.getOpcode() != AArch64::FORM_TRANSPOSED_REG_TUPLE_X4_PSEUDO) |
| 1223 | continue; |
| 1224 | |
| 1225 | unsigned UseOps = Use.getNumOperands() - 1; |
| 1226 | const TargetRegisterClass *StridedRC; |
| 1227 | switch (RegID) { |
| 1228 | case AArch64::ZPR2StridedOrContiguousRegClassID: |
| 1229 | StridedRC = &AArch64::ZPR2StridedRegClass; |
| 1230 | break; |
| 1231 | case AArch64::ZPR4StridedOrContiguousRegClassID: |
| 1232 | StridedRC = &AArch64::ZPR4StridedRegClass; |
| 1233 | break; |
| 1234 | default: |
| 1235 | llvm_unreachable("Unexpected RegID" ); |
| 1236 | } |
| 1237 | |
| 1238 | SmallVector<MCPhysReg, 4> StridedOrder; |
| 1239 | for (MCPhysReg Reg : Order) |
| 1240 | if (StridedRC->contains(Reg)) |
| 1241 | StridedOrder.push_back(Elt: Reg); |
| 1242 | |
| 1243 | int OpIdx = Use.findRegisterUseOperandIdx(Reg: VirtReg, TRI: this); |
| 1244 | assert(OpIdx != -1 && "Expected operand index from register use." ); |
| 1245 | |
| 1246 | unsigned TupleID = MRI.getRegClass(Reg: Use.getOperand(i: 0).getReg())->getID(); |
| 1247 | bool IsMulZPR = TupleID == AArch64::ZPR2Mul2RegClassID || |
| 1248 | TupleID == AArch64::ZPR4Mul4RegClassID; |
| 1249 | |
| 1250 | const MachineOperand *AssignedRegOp = llvm::find_if( |
| 1251 | Range: make_range(x: Use.operands_begin() + 1, y: Use.operands_end()), |
| 1252 | P: [&VRM](const MachineOperand &Op) { |
| 1253 | return VRM->hasPhys(virtReg: Op.getReg()); |
| 1254 | }); |
| 1255 | |
| 1256 | // Example: |
| 1257 | // |
| 1258 | // When trying to find a suitable register allocation for VirtReg %v2 in: |
| 1259 | // |
| 1260 | // %v0:zpr2stridedorcontiguous = ld1 p0/z, [...] |
| 1261 | // %v1:zpr2stridedorcontiguous = ld1 p0/z, [...] |
| 1262 | // %v2:zpr2stridedorcontiguous = ld1 p0/z, [...] |
| 1263 | // %v3:zpr2stridedorcontiguous = ld1 p0/z, [...] |
| 1264 | // %v4:zpr4mul4 = FORM_TRANSPOSED_X4 %v0:0, %v1:0, %v2:0, %v3:0 |
| 1265 | // |
| 1266 | // One such suitable allocation would be: |
| 1267 | // |
| 1268 | // { z0, z8 } = ld1 p0/z, [...] |
| 1269 | // { z1, z9 } = ld1 p0/z, [...] |
| 1270 | // { z2, z10 } = ld1 p0/z, [...] |
| 1271 | // { z3, z11 } = ld1 p0/z, [...] |
| 1272 | // { z0, z1, z2, z3 } = |
| 1273 | // FORM_TRANSPOSED_X4 {z0, z8}:0, {z1, z9}:0, {z2, z10}:0, {z3, z11}:0 |
| 1274 | // |
| 1275 | // Below we distinguish two cases when trying to find a register: |
| 1276 | // * None of the registers used by FORM_TRANSPOSED_X4 have been assigned |
| 1277 | // yet. In this case the code muse ensure that there are at least UseOps |
| 1278 | // free consecutive registers. If IsMulZPR is true, then the first of |
| 1279 | // registers must also be a multiple of UseOps, e.g. { z0, z1, z2, z3 } |
| 1280 | // is valid but { z1, z2, z3, z5 } is not. |
| 1281 | // * One or more of the registers used by FORM_TRANSPOSED_X4 is already |
| 1282 | // assigned a physical register, which means only checking that a |
| 1283 | // consecutive range of free tuple registers exists which includes |
| 1284 | // the assigned register. |
| 1285 | // e.g. in the example above, if { z0, z8 } is already allocated for |
| 1286 | // %v0, we just need to ensure that { z1, z9 }, { z2, z10 } and |
| 1287 | // { z3, z11 } are also free. If so, we add { z2, z10 }. |
| 1288 | |
| 1289 | if (AssignedRegOp == Use.operands_end()) { |
| 1290 | // There are no registers already assigned to any of the pseudo |
| 1291 | // operands. Look for a valid starting register for the group. |
| 1292 | for (unsigned I = 0; I < StridedOrder.size(); ++I) { |
| 1293 | MCPhysReg Reg = StridedOrder[I]; |
| 1294 | |
| 1295 | // If the FORM_TRANSPOSE nodes use the ZPRMul classes, the starting |
| 1296 | // register of the first load should be a multiple of 2 or 4. |
| 1297 | unsigned SubRegIdx = Use.getOperand(i: OpIdx).getSubReg(); |
| 1298 | if (IsMulZPR && (getSubReg(Reg, Idx: SubRegIdx) - AArch64::Z0) % UseOps != |
| 1299 | ((unsigned)OpIdx - 1)) |
| 1300 | continue; |
| 1301 | |
| 1302 | // In the example above, if VirtReg is the third operand of the |
| 1303 | // tuple (%v2) and Reg == Z2_Z10, then we need to make sure that |
| 1304 | // Z0_Z8, Z1_Z9 and Z3_Z11 are also available. |
| 1305 | auto IsFreeConsecutiveReg = [&](unsigned UseOp) { |
| 1306 | unsigned R = Reg - (OpIdx - 1) + UseOp; |
| 1307 | return StridedRC->contains(Reg: R) && |
| 1308 | (UseOp == 0 || |
| 1309 | ((getSubReg(Reg: R, Idx: AArch64::zsub0) - AArch64::Z0) == |
| 1310 | (getSubReg(Reg: R - 1, Idx: AArch64::zsub0) - AArch64::Z0) + 1)) && |
| 1311 | !Matrix->isPhysRegUsed(PhysReg: R); |
| 1312 | }; |
| 1313 | if (all_of(Range: iota_range<unsigned>(0U, UseOps, /*Inclusive=*/false), |
| 1314 | P: IsFreeConsecutiveReg)) |
| 1315 | Hints.push_back(Elt: Reg); |
| 1316 | } |
| 1317 | } else { |
| 1318 | // At least one operand already has a physical register assigned. |
| 1319 | // Find the starting sub-register of this and use it to work out the |
| 1320 | // correct strided register to suggest based on the current op index. |
| 1321 | MCPhysReg TargetStartReg = |
| 1322 | getSubReg(Reg: VRM->getPhys(virtReg: AssignedRegOp->getReg()), Idx: AArch64::zsub0) + |
| 1323 | (OpIdx - AssignedRegOp->getOperandNo()); |
| 1324 | |
| 1325 | for (unsigned I = 0; I < StridedOrder.size(); ++I) |
| 1326 | if (getSubReg(Reg: StridedOrder[I], Idx: AArch64::zsub0) == TargetStartReg) |
| 1327 | Hints.push_back(Elt: StridedOrder[I]); |
| 1328 | } |
| 1329 | |
| 1330 | if (!Hints.empty()) |
| 1331 | return TargetRegisterInfo::getRegAllocationHints(VirtReg, Order, Hints, |
| 1332 | MF, VRM); |
| 1333 | } |
| 1334 | } |
| 1335 | |
| 1336 | for (MachineInstr &MI : MRI.def_instructions(Reg: VirtReg)) { |
| 1337 | if (MI.getOpcode() != AArch64::FORM_TRANSPOSED_REG_TUPLE_X2_PSEUDO && |
| 1338 | MI.getOpcode() != AArch64::FORM_TRANSPOSED_REG_TUPLE_X4_PSEUDO) |
| 1339 | return TargetRegisterInfo::getRegAllocationHints(VirtReg, Order, Hints, |
| 1340 | MF, VRM); |
| 1341 | |
| 1342 | unsigned FirstOpSubReg = MI.getOperand(i: 1).getSubReg(); |
| 1343 | switch (FirstOpSubReg) { |
| 1344 | case AArch64::zsub0: |
| 1345 | case AArch64::zsub1: |
| 1346 | case AArch64::zsub2: |
| 1347 | case AArch64::zsub3: |
| 1348 | break; |
| 1349 | default: |
| 1350 | continue; |
| 1351 | } |
| 1352 | |
| 1353 | // Look up the physical register mapped to the first operand of the pseudo. |
| 1354 | Register FirstOpVirtReg = MI.getOperand(i: 1).getReg(); |
| 1355 | if (!VRM->hasPhys(virtReg: FirstOpVirtReg)) |
| 1356 | continue; |
| 1357 | |
| 1358 | MCRegister TupleStartReg = |
| 1359 | getSubReg(Reg: VRM->getPhys(virtReg: FirstOpVirtReg), Idx: FirstOpSubReg); |
| 1360 | for (unsigned I = 0; I < Order.size(); ++I) |
| 1361 | if (MCRegister R = getSubReg(Reg: Order[I], Idx: AArch64::zsub0)) |
| 1362 | if (R == TupleStartReg) |
| 1363 | Hints.push_back(Elt: Order[I]); |
| 1364 | } |
| 1365 | |
| 1366 | return TargetRegisterInfo::getRegAllocationHints(VirtReg, Order, Hints, MF, |
| 1367 | VRM); |
| 1368 | } |
| 1369 | |
| 1370 | unsigned AArch64RegisterInfo::getLocalAddressRegister( |
| 1371 | const MachineFunction &MF) const { |
| 1372 | const auto &MFI = MF.getFrameInfo(); |
| 1373 | if (!MF.hasEHFunclets() && !MFI.hasVarSizedObjects()) |
| 1374 | return AArch64::SP; |
| 1375 | else if (hasStackRealignment(MF)) |
| 1376 | return getBaseRegister(); |
| 1377 | return getFrameRegister(MF); |
| 1378 | } |
| 1379 | |
| 1380 | /// SrcRC and DstRC will be morphed into NewRC if this returns true |
| 1381 | bool AArch64RegisterInfo::shouldCoalesce( |
| 1382 | MachineInstr *MI, const TargetRegisterClass *SrcRC, unsigned SubReg, |
| 1383 | const TargetRegisterClass *DstRC, unsigned DstSubReg, |
| 1384 | const TargetRegisterClass *NewRC, LiveIntervals &LIS) const { |
| 1385 | MachineFunction &MF = *MI->getMF(); |
| 1386 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
| 1387 | |
| 1388 | if (MI->isSubregToReg() && MRI.subRegLivenessEnabled() && |
| 1389 | !MF.getSubtarget<AArch64Subtarget>().enableSRLTSubregToRegMitigation()) |
| 1390 | return false; |
| 1391 | |
| 1392 | if (MI->isCopy() && |
| 1393 | ((DstRC->getID() == AArch64::GPR64RegClassID) || |
| 1394 | (DstRC->getID() == AArch64::GPR64commonRegClassID)) && |
| 1395 | MI->getOperand(i: 0).getSubReg() && MI->getOperand(i: 1).getSubReg()) |
| 1396 | // Do not coalesce in the case of a 32-bit subregister copy |
| 1397 | // which implements a 32 to 64 bit zero extension |
| 1398 | // which relies on the upper 32 bits being zeroed. |
| 1399 | return false; |
| 1400 | |
| 1401 | auto IsCoalescerBarrier = [](const MachineInstr &MI) { |
| 1402 | switch (MI.getOpcode()) { |
| 1403 | case AArch64::COALESCER_BARRIER_FPR16: |
| 1404 | case AArch64::COALESCER_BARRIER_FPR32: |
| 1405 | case AArch64::COALESCER_BARRIER_FPR64: |
| 1406 | case AArch64::COALESCER_BARRIER_FPR128: |
| 1407 | return true; |
| 1408 | default: |
| 1409 | return false; |
| 1410 | } |
| 1411 | }; |
| 1412 | |
| 1413 | // For calls that temporarily have to toggle streaming mode as part of the |
| 1414 | // call-sequence, we need to be more careful when coalescing copy instructions |
| 1415 | // so that we don't end up coalescing the NEON/FP result or argument register |
| 1416 | // with a whole Z-register, such that after coalescing the register allocator |
| 1417 | // will try to spill/reload the entire Z register. |
| 1418 | // |
| 1419 | // We do this by checking if the node has any defs/uses that are |
| 1420 | // COALESCER_BARRIER pseudos. These are 'nops' in practice, but they exist to |
| 1421 | // instruct the coalescer to avoid coalescing the copy. |
| 1422 | if (MI->isCopy() && SubReg != DstSubReg && |
| 1423 | (AArch64::ZPRRegClass.hasSubClassEq(RC: DstRC) || |
| 1424 | AArch64::ZPRRegClass.hasSubClassEq(RC: SrcRC))) { |
| 1425 | unsigned SrcReg = MI->getOperand(i: 1).getReg(); |
| 1426 | if (any_of(Range: MRI.def_instructions(Reg: SrcReg), P: IsCoalescerBarrier)) |
| 1427 | return false; |
| 1428 | unsigned DstReg = MI->getOperand(i: 0).getReg(); |
| 1429 | if (any_of(Range: MRI.use_nodbg_instructions(Reg: DstReg), P: IsCoalescerBarrier)) |
| 1430 | return false; |
| 1431 | } |
| 1432 | |
| 1433 | return true; |
| 1434 | } |
| 1435 | |
| 1436 | bool AArch64RegisterInfo::shouldAnalyzePhysregInMachineLoopInfo( |
| 1437 | MCRegister R) const { |
| 1438 | return R == AArch64::VG; |
| 1439 | } |
| 1440 | |
| 1441 | bool AArch64RegisterInfo::isIgnoredCVReg(MCRegister LLVMReg) const { |
| 1442 | return (LLVMReg >= AArch64::Z0 && LLVMReg <= AArch64::Z31) || |
| 1443 | (LLVMReg >= AArch64::P0 && LLVMReg <= AArch64::P15); |
| 1444 | } |
| 1445 | |